Amazon Machine Learning

Overview

• Original Machine learning service
• In 2017 Amazon SageMaker released

"Model"

• Supervised machine learning
  • Training data (label: spam/not-spam) - max 100GB
  • Tuned until receives desired accuracy

Data

• Input: RDS, S3 (CSV), Redshift
• Output: S3

Training Data

• Historical data
• Contain result (Target attribute)
• Example
  • transaction details: spam/not-spam

Process

• Build model
  • Create datasource
  • Explore and understand your data
    • ML computes the statistics
  • Create a model
    
    • Select data source
    • Model type
      
      • BINARY CLASSIFICATION
        • Yes/No
      • REGRESSION
        • Predicts a number
        • e.g. how much will this house sell for
      • MULTI CLASSIFICATION
        • Assign a category (e.g. genre) 
    • Each model type has evaluation score
    • Reciple
      • Transformations applied to variables
• Evaluate and optimize
  • All model types have visualization
  • You can tweak parameters
• Retrieve predictions
  • Batch: large volume prediction analysis
    • Async
    • They are output to S3
  • Real-Time: 
    • Sync
    • Low-latency

Use cases

• Predictions
  • Will this customer buy the product
  • Is this order fraudulent
• Recommendations
  • What other articles are interesting
• Targeted marketing
• Content classifications

Notes

• Two modes
  
  • Interactive (experimentation)
  • API (automated access)
    • Batch predictions
    • Real-Time predictions